Abstract:
In this paper, a novel approach for initializing the spherical K-means algorithm is proposed.
It is based on calculating well distributed seeds across the input space. Also, a new measure
for calculating vectors? directional variance is formulated, to be used as a measure of clusters?
compactness. The proposed initialization scheme is compared with the classical Kmeans
? where initial seeds are specified randomly or arbitrarily ? on two datasets. The
assessment was based on three measures: an objective function that measures intra cluster
similarity, cluster compactness and time to converge. The proposed algorithm (called initialized
K-means) outperforms the classical (random) K-means when intra cluster similarity
and cluster compactness were considered for several values of k (number of clusters). As
far as convergence time is concerned, the initialized K-means converges faster than the
random K-means for small number of clusters. For a large number of clusters the time
necessary to calculate the initial clusters? seeds start to outweigh the convergence criterion
in time. The exact number of clusters at which the proposed algorithm starts to change
behavior is data dependent (=11 for dataset1 and = 15 for dataset2).